Experimental Results

We conducted a series of experiments with real usage data from the site for the newsletter of the Association for Consumer Research (from July 1998 to June 1999). The site contains a variety of news items, including President's columns, conference announcements, and call-for-papers for a number of conferences and journals. The usage preprocessing steps described earlier resulted in a user transaction file containing 18430 user transactions with a total of 62 pageviews represented uniquely by their associated URLs. The transaction clustering process yielded 16 transaction clusters representing different types of user access patterns. A threshold of 0.5 was used to derive usage profiles from transaction clusters (i.e., profiles contained only those pageviews appearing in at least 50% of transactions). In the content preprocessing stage, a total of 566 significant features were extracted with each document contributing at most 20 significant features to the global dictionary (normalized term frequency was used for measuring feature significance). Feature clustering using multivariate k-means resulted in 28 feature clusters from which the corresponding (overlapping) content profiles were derived.

Figure 2 depicts an example of two overlapping content profiles. The top significant features where listed for each document for illustrative purposes. These features indicated why each of the pageviews were included in each content profile. The first profile captures those documents in which a portion of the content relates to global and international business management and marketing. On the other hand, the second profile includes documents about consumer behavior and psychology in marketing. Note that documents which contain content related to both topics have been included in both profiles. Usage profiles are represented in the same manner, but they capture overlapping aggregate usage patterns of the site users.

  

Figure 2: Two Overlapping Content Profiles

The recommendation engine was used for a sample user session using a window size of 2. Figure 3 shows the system recommendations based only on usage profiles, while Figure 2 shows the results from only the content profiles. In these tables, the first column shows the pageviews contained in the current active session. The last pageview in each session window represents the current location of the user in the site. The right-hand column gives the recommendation score obtained using the techniques discussed in the previous section. It is clear from these examples that the combination of recommendations from both content and usage profiles can provide added value to the user. For example, in the usage-based recommendations, the user's visit to "ACR Board of Directors Meeting" did not yield any recommendations with a score above the specified threshold (0.5), while content-based recommendations produced some pages with related content. On the other hand, navigating to the page "Conference Update" resulted in content-based recommendations for pages with only general information and news about conferences, while usage profiles yielded a number of specific recommendations that the site users interested in conferences and calls for papers tend visit.

  

Figure 3: Recommendations Based on Usage Profiles

  

Figure 4: Recommendations Based on Content Profiles